Speaker
Description
Scientific data compression has become a critical component in large-scale computing and experimental facilities, where data volumes increasingly outpace storage capacity and I/O bandwidth. This report surveys recent advances and the current state of the field. We review the evolution of error-bounded lossy compression algorithms, including SZ, ZFP, and MGARD, with emphasis on their extended capabilities for feature-preserving and physics-aware reduction. We examine the emerging paradigm of compressed-domain computation, which enables analytic operations to be executed directly on compressed representations without full reconstruction. Hardware acceleration support, including SIMD instructions and offloading to DPUs, is summarized alongside ongoing standardization efforts. The report also addresses the evaluation of scientific fidelity—specifically, methodologies for quantifying compression-induced distortion in downstream analyses. We conclude with a discussion of remaining limitations, including reproducibility concerns, heterogeneous platform portability, and the integration of machine learning techniques into compression pipelines.
